Model calibration audit

For every bucket of predicted probability, what did the model actually hit? A well-calibrated model has actual win rate ≈ predicted prob. A negative gap means the model is overconfident in that bucket (dangerous); positive gap means it's underpredicting (safe). Rows turn red when |gap| > 5pp AND N ≥ 10.

Calibration cohort by sport (click to filter):

All sports MLB N=3,353 TENNIS_WTA N=753 TENNIS_ATP N=748 MMA_MIXED_MARTIAL_ARTS N=407 LIGUE1 N=89 NFL_PRESEASON N=65 NBA N=42 NHL N=33 LALIGA N=26 SERIEA N=19 EPL N=15 UCL N=13 NCAAB N=11 CRICKET_IPL N=6 BUNDESLIGA N=3

Combined "all sports" is rarely meaningful — sports differ in market efficiency, signal availability, and base rates. Use the chips to drill into a single sport. N<50 (red) means the calibration is brittle; N≥200 (green) is trustworthy.

Filter: window=90d · sport=nfl_preseason

Overall: N = 65 · mean predicted 55.7% · actual win rate 58.5% · gap +2.8pp · Brier 0.246 · log-loss 0.686

Calibration by predicted-probability bucket.
Predicted-prob bucket	N	Mean predicted	Actual win rate	Gap (actual − predicted)	Brier
<50%	7	49.6%	42.9%	-6.7pp	0.251
50-55%	28	52.5%	64.3%	+11.8pp	0.247
55-60%	19	57.3%	52.6%	-4.7pp	0.250
60-65%	8	61.5%	62.5%	+1.0pp	0.242
70-75%	3	73.4%	66.7%	-6.7pp	0.224

JSON feed · Model CLV · Methodology

Calibration is computed from the prediction_history table. Bucket edges are 5pp wide (50-55%, 55-60%, etc.) with widened tails (<50%, 80-90%, 90%+) where data is sparser. Negative gap is the dangerous direction — actual win rate below stated confidence.